TabLOID

A model learning to browse does not necessarily need the keys to the live web. In an October 23 update, OpenAI says frontier training now uses cached, indexed web material inside an offline environment.

The policy concerns internal research, including models still undergoing safety work. It is not a shutdown of internet access for released customer products.

OpenAI says a September research incident exposed a route to limited internet access. Its response combines stronger isolation, monitoring and deliberate attempts to break its own defenses. The surrounding infrastructure is not completely disconnected: running experiments still requires external connections.

The interesting engineering problem is keeping the boundary intact even when the thing inside is clever enough to look for a loophole. “Please behave” is a short security manual.

OpenAI’s account of its training safeguards

Anthropic’s October 15 account describes Claude review on every pull request, followed by security checks, tests and human approvals scaled to risk. Some low-risk changes from trusted authors can clear without an additional person; sensitive work still needs people. An agent’s approval cannot substitute for a required human one.

Uber’s October 20 engineering post tackles another chore: deleting obsolete feature-flag code. Its AutoClean system combines deterministic rewriting tools with agents, then runs tests, static checks and four focused verification agents. Only changes that clear those checks reach a human reviewer.

These are the companies’ descriptions of their own systems, not independent evidence that automated review catches everything. The shared lesson is practical: a machine producing more work needs a sharper inspection process. A green test result can still leave a very interesting mistake.

Anthropic’s review process
Uber’s code-cleanup system

In a study published September 23 in Aging, researchers reduced the magnetic field around male fruit flies. Flies with a loss-of-function mutation in Pink1, used as a model of neurodegeneration, lived about 20% longer under the shielded conditions.

Their climbing ability got worse. Healthy control flies improved at climbing but did not gain the same lifespan benefit.

That split is the story. A longer life and better physical function are different outcomes, even in a small insect. The researchers also measured changes in mitochondrial activity, the machinery cells use to manage energy.

This was an animal experiment under specific laboratory conditions. It does not establish a way for people to live longer, and it offers no reason to turn your bedroom into a magnetic bunker. Biology has once again declined the tidy headline.

Read the original fly study

Knowing where a pendulum is does not tell you where it is going. Direction and speed matter too.

An October 11 preprint introduces LeHami, a model that builds position-and-momentum structure into its internal representation. On visual pendulum tasks, the authors report stronger velocity predictions and more stable long-horizon dynamics than their unconstrained comparison model.

The boundaries matter. These are pendulum experiments, not a demonstration of general understanding of the physical world. The comparison model did better on some position-focused tasks, and extending the approach to strongly dissipative systems, such as heavy friction, remains future work.

It is a useful research question: which rules should a model learn, and which should its design refuse to let it forget? Even an ambitious world model can start by getting the swing right.

Read the LeHami preprint